Search

Data representation forms a feature space where forms data distribution that is one of the key factors determining the prediction accuracy of machine learning (ML). In particular, the data representation is crucial to handle small and biased training datasets, which is the main challenge of ML in chemical applications. In this paper, we propose a data-agnostic representation method that automatically and universally generates a vector-shaped and target-specified representation of crystal structures. By employing the new materials representation of the proposed method, the prediction capabilities of ML algorithms were highly improved on small training datasets and transfer learning tasks. Moreover, the prediction accuracies of ML algorithms were improved by 28.89-30.87% in extrapolation problems to predict the physical properties of the materials in unknown material groups. The source code of EMRL is publicly available at https://github.com/ngs00/emrl/tree/master/EMRL.

Feature extraction has been widely studied to find informative latent features and reduce the dimensionality of data. In particular, due to the difficulty in obtaining labeled data, unsupervised feature extraction has received much attention in data mining. However, widely used unsupervised feature extraction methods require side information about data or rigid assumptions on the latent feature space. Furthermore, most feature extraction methods require predefined dimensionality of the latent feature space,which should be manually tuned as a hyperparameter. In this article, we propose a new unsupervised feature extraction method called Unsupervised Subspace Extractor ( USE ), which does not require any side information and rigid assumptions on data. Furthermore, USE can find a subspace generated by a nonlinear combination of the input feature and automatically determine the optimal dimensionality of the subspace for the given nonlinear combination. The feature extraction process of USE is well justified mathematically, and we also empirically demonstrate the effectiveness of USE for several benchmark datasets.

Chemical reactions typically have numerous controllable factors that need to be optimized to yield the desired products. Although traditional experimental methods are limited to explore possible combinations of these factors, artificial intelligence (AI) can provide the optimal solution based on chemical reaction data. In this study, we optimize the non-oxidative conversion of methane to C2 compounds using AI, such as machine learning (ML) to predict experimental results and metaheuristics to optimize reaction conditions. A decision tree-based machine learning method can reasonably predict the reaction outcomes (CH4 conversion, C2 yield, and selectivities for C2 and coke) with an error of < 5%. Trained ML models are applied to maximize the C2 yield by optimizing the reaction parameters with metaheuristics. We can simultaneously enhance the C2 yield and suppress the coke formation by improving the multi-objective function for the optimization. We believe that our method will be helpful to optimize the chemical reaction conditions with multiple targets.

The open-access material databases allowed us to approach scientific questions from a completely new perspective with machine learning methods. Here, on the basis of open-access databases, we focus on the classical band gap problem for predicting accurately the band gap of a crystalline compound using a machine learning approach with newly developed tuplewise graph neural networks (TGNN), which is devised to automatically generate input representation of crystal structures in tuple types and to exploit crystal-level properties as one of the input features. Our method brings about a highly accurate prediction of the band gaps at hybrid functionals and GW approximation levels for multiple material data sets without heavy computational cost. Furthermore, to demonstrate the applicability of our prediction model, we provide a data set of GW band gaps for 45835 materials predicted by TGNN posing higher accuracy than standard density functional theory calculations.

The fundamental goal of machine learning (ML) in physical science is to predict the physical properties of unobserved states. However, an accurate predictionfor input data outside of training distributions is a challenging problem in ML due to nonlinearities of input and target dynamics. For an accurate extrapolation of ML algorithms, we propose a new data-driven method that encodes nonlinearities of physical systems into input representations. Based on the proposed encoder, a given physical system is described as linear-like functions that are easy to extrapolate. By applying the proposed encoder, the extrapolation errors were significantly reduced by 48.39% and 40.04% in an n-body problem and a prediction of materials properties, respectively.

With great growth in the volume of chemical databases, machine learning receives significant attention from various scientific communities for efficient high-throughput screening of molecular properties and drug discovery on the millions of chemical compounds. In particular, graph neural networks (GNNs) have been widely studied in chemistry-related fields because a molecule is natively represented as a mathematical graph. In GNNs for the graph-level analysis, a global operation called readout is applied after node embedding to generate a graph-level embedding that represents characteristics of the whole graph. However, commonly used readouts frequently distort scale information of the graph and consequently degrade the prediction accuracy of GNNs. This problem becomes more serious in molecular machine learning because molecules have many important scale information (e.g., molecular weight and total energy). In this paper, we investigate this scale distortion problem in GNNs caused by the readouts for the first time and propose an efficient solution with a new attention-based readout. In the experiments, the proposed readout outperformed commonly used readouts on various GNN architectures.

Estimating the importance of each atom in a molecule is one of the most appealing and challenging problems in chemistry, physics, and materials science. The most common way to estimate the atomic importance is to compute the electronic structure using density functional theory (DFT), and then to interpret it using domain knowledge of human experts. However, this conventional approach is impractical to the large molecular database because DFT calculation requires large computation, specifically, O(n4) time complexity w.r.t. the number of electronic basis functions. Furthermore, the calculation results should be manually interpreted by human experts to estimate the atomic importance in terms of the target molecular property. To tackle this problem, we first exploit the machine learning-based approach for the atomic importance estimation based on the reverse self-attention on graph neural networks and integrating it with graph-based molecular description. Our method provides an efficiently-automated and target-directed way to estimate the atomic importance without any domain knowledge of chemistry and physics.

With precipitously growing demand to detect outliers in data streams, many studies have been conducted aiming to develop extensions of well-known outlier detection algorithm called Local Outlier Factor (LOF), for data streams. However, existing LOF-based algorithms for data streams still suffer from two inherent limitations: 1) Large amount of memory space is required. 2) A long sequence of outliers is not detected. In this paper, we propose a new outlier detection algorithm for data streams, called DILOF that effectively overcomes the limitations. To this end, we first develop a novel density-based sampling algorithm to summarize past data and then propose a new strategy for detecting a sequence of outliers. It is worth noting that our proposing algorithms do not require any prior knowledge or assumptions on data distribution. Moreover, we accelerate the execution time of DILOF about 15 times by developing a powerful distance approximation technique. Our comprehensive experiments on real-world datasets demonstrate that DILOF significantly outperforms the state-of-the-art competitors in terms of accuracy and execution time. The source code for the proposed algorithm is available at our website: http://di.postech.ac.kr/DILOF.